Deprecate Allocate-Load

[blegat]

The purpose of Allocate-Load is to be able to communicate the model in two-pass across layers without the need to create intermediate CachingOptimizer.
However, it's not yet implemented by the bridge optimizers and it's not clear to me whether there is a way to do this without complicating the job of creating new bridges for users.
This new features #1286 allows to pass non-bridges constraints across layers which allows to avoid copying the function as the destination model of the copy can get the CanonicalConstraintFunction directly from the source model.
So a model needed Allocate-Load could implement MOIU.pass_nonvariable_constraints instead and it would work if no constraint if bridges.
To also work with bridging constraints, it should also implement MOI.add_constraint but only allow it to be called during the copy operation.
This is what is done for MatrixOfConstraints: #1287.
In fact, I expect most solver needing Allocate-Load to be rewritten with MatrixOfConstraints like Clp in jump-dev/Clp.jl#111.
The only two exceptions are CSDP and SDPA but there is no point in keeping the Allocate-Load API just for these two.
I'll keep this open until we're close to releasing v0.10 to evaluate if things happens as expected and then we can decide.

[odow]

There are a few major benefits of getting rid of allocate-load

• Code simplicity. Currently, there is a whole file of code that isn't really used, and just complicates the whole copy story. We don't need three ways of implementing copy!
• Getting rid of the pass_attrs and pass_cons to the copy functions. These really reduce the readability of the copy operations and add to the strain on inference.

I still don't really understand the need for MOIU.pass_nonvariable_constraints. If solvers have unique copy requirements concerning the order of variables, can't they just write a new copy_to? We don't want this copy being too generic that solvers can patch different parts of it.

[blegat]

I don't know if many solvers are going to use MOIU.pass_nonvariable_constraints. The main use-case is MatrixOfConstraints.

[odow]

Time to revisit this @blegat?

[blegat]

When updating the solvers using it for MOI v0.10, I'll try at the same time to stop using allocate-load and verify it's as expected.

[odow]

@blegat do you have time at some point to try migrating one of the allocate-load solvers? This is one of the blockers for MOI 0.10.

[blegat]

Updating the conic solvers is the next item on my TODO list after #1461